The invention relates to a personalized commodity recommendation method. The method comprises the steps of receiving a commodity search instruction of a user; If the users are classified, calculatingmulti-dimensional average label information preferred by the users, wherein a recommendation model of the users calculates first multi-dimensional attribute information based on the multi-dimensionalaverage label information, calculates the distance between the first multi-dimensional attribute information and second multi-dimensional attribute information of each commodity in the commodity category, and pushes the commodities with the smaller distance to the users; Otherwise, calculating first multi-dimensional average attribute information based on the second multi-dimensional attribute information of the plurality of commodities, and determining a recommendation model suitable for the user based on the first multi-dimensional average attribute information and the recommendation model of each type of users. According to the recommendation method provided by the invention, firstly, users are clustered, and multi-dimensional commodity label information and commodity attribute information are utilized, so that the recommendation accuracy is high; Besides, when the user is a new user, a recommendation model of the new user is formed by utilizing an existing recommendation model through attention transfer learning, so that the cold start problem is solved.